All research systems
Agentic FabricActive

What does an orchestration layer owe the guarantees below it?

Question → hypothesis/design → architecture → evaluation → results → limitations → next questions

Motivation

Prompts ask an agent to follow a process; nothing checks that it did. Agentic Fabric publishes the patterns, workflows and decisions learned from running Claude Code agents daily, and ships the parts a hook can enforce as plugins, each with the tests that try to break it.

Hypothesis

Rules that matter in an agent's workflow hold only when they are hooks the agent cannot argue past, and a security hook holds only when it accepts one form it can fully parse and refuses every other.

Architecture

  • Patterns and anti-patterns from operating agents, each labelled with whether its evidence is reproduced by a test in the repository or observed in operation
  • Workflows: review-first pull requests, cross-vendor review with bounded cycles, turning escaped defects into eval cases, measuring from existing logs before adding instrumentation
  • fabric-gates, a Claude Code plugin: draft-first pull requests, merges only when GitHub's head matches the branch tip, a branch prefix, and edits to the enforcement layer only after a human unlock
  • Decision records that keep the options they rejected

Evaluation methodology

The pull request gates have shell regression suites that keep the bypasses found in review as test cases, plus a script that reproduces the fail-closed behaviour. Patterns label their evidence as reproduced or observed.

Results

Preliminary results

Limitations

  • The gates see the agent's tool calls, not its intent; a shell alias, script or the web interface bypasses them, so server-side branch rules stay on
  • Observed evidence comes from one maintainer's operation and says so

Current status

v0.1.0 is published: patterns, workflows, decision records and the fabric-gates plugin. Model routing, review loops with a seeded-defect eval, workflow metrics and engineering standards are planned as further plugins.

Roadmap

  • Model routing plugin
  • Review plugin with the seeded-defect eval
  • Workflow metrics plugin

Next questions

  • Composition semantics for enforced routines
  • Failure containment across multi-agent routines